DATA / CHALLENGE INPUTS

Choose your starting point.

Training files are readable in advance for transfer and environment checks; challenge analysis and modeling start at kickoff. Test inputs remain encrypted until 13:30 German time on Saturday 24 October. Technical pages are available now. The participant guide and separate starter code open at 10:00 German time on event day. Modeling starts at 11:00.

The datasets we provide

Component Size Use
Modeling tables 76.1 MB Core labels, 289 numeric predictors across four sources, fixed splits, optional CV and practice requests. Start here.
Bird exploration 312.5 MB 2,490,462 German records, all available months, 492 species keys and 15 datasets.
German patches See manifest Both scales for 38,374 locations; 210 compressed NPZ shards, masks and index.
Optional German rasters 148.57 GB 87 native GeoTIFFs clipped with a uniform Germany-plus-2 km guard.
Encrypted competition inputs See manifest 6,359 anonymous locations, both patch scales and compact tables. Preload before 13:30 German time.

The complete distribution is about 188.40 GB. A table-only final model needs the 11.2 MB encrypted table archive; optional test patches add 5.52 GB. Keep repeatedly accessed files on the runtime’s local disk rather than reading through mounted Drive.

Final test-data unlock key

Opens on Saturday 24 October at 13:30 CEST. Preload the encrypted inputs and save the released age identity file for the starter unlock helper.

Open test unlock key →

Google Drive opens in a separate tab. If your browser blocks it, copy this address and open it directly in your browser:

https://drive.google.com/drive/folders/1gXsJyZnI8QIbk1RQYL0qNqy1yf0eSaoT

Participant guide · opens at 10:00 · Read the practical guide and FAQ →

What an environmental location contains

Compact table

A native 10 m embedding vector, circular summaries over 100 m and 1,000 m radii, and three coverage fractions. AlphaEarth contributes 195 predictors; WorldCover adds 39, terrain 19 and historical OSM 36, for 289 in total. The starter joins the companion tables automatically. Context summaries are not single coarse pixels. WorldCover is from 2021 and OSM from January 2024; retain their separate notices.

Local habitat

64 channels × 128 × 128 pixels at 10 m. A 1.28 km footprint, with north-up axes, validity masks and focal-cell offsets.

Wider context

64 channels × 64 × 64 pixels at 40 m. A 2.56 km footprint using Google’s normalized overview. Never average encoded int8 values directly.

Decode valid int8 values using sign(q) × (q / 127.5)²; −128 means NoData. Follow the packaged spatial schema for normalization, masks, axes, focal cells and offsets. National GeoTIFFs retain source CRS and south-up transforms; the released patches are north-up.

Plan final inference around the supplied windows. Both patch scales are available for every final location as compressed NPZ arrays, with the same channels, masks and focal offsets as the German patches. Final inputs have anonymous IDs and relative offsets; continuous georeferenced test-region GeoTIFFs are not supplied. A model trained from Germany’s whole rasters must therefore compute its final features from these windows. Check this before choosing features that require a larger area or absolute coordinates.

Ten scored targets

Species Scientific name German name
Eurasian Skylark Alauda arvensis Feldlerche
Yellowhammer Emberiza citrinella Goldammer
Common Linnet Linaria cannabina Bluthänfling
Red-backed Shrike Lanius collurio Neuntöter
Northern Lapwing Vanellus vanellus Kiebitz
Eurasian Nuthatch Sitta europaea Kleiber
Common Reed Warbler Acrocephalus scirpaceus Teichrohrsänger
Common Cuckoo Cuculus canorus Kuckuck
Whinchat Saxicola rubetra Braunkehlchen
Lesser Spotted Woodpecker Dryobates minor Kleinspecht

Use the stable species IDs and seasonal windows in species.csv. The exploration archive includes all available months; the modeling labels use the frozen species-specific windows.

Labels and validation

The focal tables contain 83,172 training labels and 19,770 German development labels. Another 74,142 training labels cover eleven helper species. Five source/block/season-matched recording-opportunity backgrounds accompany each selected positive. Missing labels remain unknown.

Everyone reports the same German development comparison. Five spatial folds are optional robustness checks; the supplied train, validation and purged roles apply across focal and helper models and all learned preprocessing.

Licenses and attribution

Bird records retain their CC0 1.0 or CC BY 4.0 terms and record/dataset credits. The GBIF source download is 10.15468/dl.wcc2ma. AlphaEarth is CC BY 4.0 with Google and Google DeepMind attribution. The German boundary mask retains BKG attribution and Datenlizenz Deutschland Namensnennung 2.0 terms. Keep the package notices when sharing derived work.

Source and attribution links do not authorize retrieving additional bird observations for this challenge. WorldCover 2021 requires its CC BY 4.0 attribution. Terrain retains the Copernicus DEM source terms and notice. The separate OSM derived tables are an ODbL 1.0 database with © OpenStreetMap contributors attribution. Keep each companion notice when copying or adapting its tables. Code licensing is separate from data licensing and participant model ownership.

Optional raster downloads

Download the 87 German GeoTIFFs from the Hugging Face raster mirror (148.57 GB in total). It contains the same raster bytes, technical notes and licenses as Challenge data. Use it if Drive returns an HTML page or a download-quota error. Download only the tiles you need and verify them against the mirror’s SHA256SUMS. Tables, patches and encrypted final inputs remain in the Drive folder.

Drive access and local copies

Challenge data is the participant download folder. It does not grant access to the parent company Drive or organizer folders. Use Viewer access and download only what your approach needs; keep repeated model reads on runtime-local storage. Do not upload presentations or outputs into this read-only data folder.

Transfer and storage checklist

  1. Choose the table or spatial route and download only required components.
  2. Verify SHA256SUMS and retain technical explanations with the datasets.
  3. Preload encrypted test files onto the actual inference runtime before 13:30 German time.
  4. Check the age tool with unrelated synthetic files before kickoff; at kickoff verify the supplied starter helper before 13:30 German time.
  5. Keep a tested table baseline as an inference fallback.